[
https://issues.apache.org/jira/browse/HBASE-15452?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15193690#comment-15193690
]
Lars Hofhansl commented on HBASE-15452:
---------------------------------------
Two runs with patch applied (test from HBASE-15453)
10 runs mean:1727.2 sigma:45.20353968440967
10 runs mean:1769.9 sigma:16.53148511174964
Two runs without patch:
10 runs mean:2194.8 sigma:56.31838065853811
10 runs mean:2183.0 sigma:31.135189095298585
So this saved about 20% of runtime! Shows again that these code paths are
extremely hot and every instruction we can save in these will be noticed!
> Consider removing checkScanOrder from StoreScanner.next
> -------------------------------------------------------
>
> Key: HBASE-15452
> URL: https://issues.apache.org/jira/browse/HBASE-15452
> Project: HBase
> Issue Type: Bug
> Reporter: Lars Hofhansl
> Attachments: 15452-0.98.txt
>
>
> In looking why we spent so much time in StoreScanner.next when doing a simple
> Phoenix count\(*) query I came across checkScanOrder. Not only is this a
> function dispatch (that the JIT would eventually inline), it also requires
> setting the prevKV member for every Cell encountered.
> Removing that logic a yields measurable end-to-end improvement of 5-20% (in
> 0.98).
> I will repeat this test on my work machine tomorrow.
> I think we're stable enough to remove that check anyway.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
